摘要
针对行人重识别研究领域中因图像背景复杂、风格变化等因素导致的ReID模型迁移能力差的问题,提出一种外观风格转换的行人数据增广方法。首先,使用循环生成对抗网络学习图像之间的不同风格;其次,使用多维细粒度语义网络对行人图像进行语义分割,为行人图像的生成提供细粒度的引导;最后,通过对外观风格与内容进行修复,完成图像整体风格与行人外观之间的转换。该方法在Market-1501与DukeMTMC-reID两种数据集上进行了仿真实验。实验结果表明:ReID模型在该方法增广的数据集上训练,切实提高了模型的迁移识别能力。
To solve the problem of poor migration ability of RelD model due to complex image background,style change and other factors in the research field of Person Re-recognition,a pedestrian data augmentation method based on appearance style con-version is proposed.First,the CycleGAN is used to learn the different style between images.Secondly,the multi-dimension fine-grained semantic network is used for semantic segmentation of pedestrian images,providing fine-grained guidance for the gen-eration of pedestrian images.Finally,by repairing the appearance style and content,the transformation between the overall image style and pedestrian appearance is completed.The method is simulated on Market-1501 and DukeMTMC-reID datasets.The experi-mental results show that the ReID model is trained on the data set augmented by the method,and the migration recognition ability of the model is effectively improved.
作者
张国辉
刘志刚
高月
ZHANG Guohui;LIU Zhigang;GAO Yue(School of Computer&Information Technology,Northeast Petroleum University,Daqing 163318;Post-doctoral Workstation on Institute of Applied Technology,Northeast Petroleum University,Daqing 163318)
出处
《计算机与数字工程》
2025年第8期2324-2328,共5页
Computer & Digital Engineering
关键词
生成对抗网络
语义分割
行人重识别
数据增广
GAN
semantic segmentation
person re-identification
data augmentation
